This study tackles the limited domain knowledge of large language models in automated instructional design, which often results in outputs lacking practicality. We propose a multi-agent collaborative framework (MATD) that simulates teacher reasoning through a two-step process: first building the overall structure, then refining key elements like teaching objectives and procedures. Using dynamic task scheduling and iterative optimization guided by structured scoring feedback, the framework enables effective collaboration among agents. Experiments show that without requiring additional training or resources, our method significantly improves the completeness and practicality of generated instructional designs, offering a new solution for automated instructional design.